no resnet.eval() in encoder ?

HHTseng / video-classification

Tutorial for video classification/ action recognition using 3D CNN/ CNN+RNN on UCF101

940 stars 216 forks source link

no resnet.eval() in encoder ? #49

Closed Gateway2745 closed 3 years ago

Gateway2745 commented 3 years ago

@HHTseng Hi! Thanks for this repo. I noticed you did not do resnet.eval() in your encoder for frame feature extraction. Since we are freezing the weights of the resnet model, does it help to use the pretrained batch norm params while normalizing with current batch stats?

https://github.com/HHTseng/video-classification/blob/99ebf204f0b1d737e38bc0d8b65aca128a57d7b1/ResNetCRNN/functions.py#L325-L341

Gateway2745 commented 3 years ago

The parameters of batch norm are not updated. However, running stats will be updated.